EE263 Autumn 2007-08 Stephen Boyd

Lecture 12
Jordan canonical form

• Jordan canonical form

• generalized modes

• Cayley-Hamilton theorem

12–1
 Jordan canonical form

what if A cannot be diagonalized?

any matrix A ∈ Rn×n can be put in Jordan canonical form by a similarity

transformation, i.e.
 
J1
T −1AT = J =  ... 
Jq

where  
λi 1
 λi ... 
Ji =   ∈ Cni×ni
 ... 1 
λi
Pq
is called a Jordan block of size ni with eigenvalue λi (so n = i=1 ni )

Jordan canonical form 12–2

• J is upper bidiagonal

• J diagonal is the special case of n Jordan blocks of size ni = 1

• Jordan form is unique (up to permutations of the blocks)

• can have multiple blocks with same eigenvalue

Jordan canonical form 12–3

note: JCF is a conceptual tool, never used in numerical computations!

X (s) = det(sI − A) = (s − λ1)n1 · · · (s − λq )nq

hence distinct eigenvalues ⇒ ni = 1 ⇒ A diagonalizable

dim N (λI − A) is the number of Jordan blocks with eigenvalue λ

more generally,
X
k
dim N (λI − A) = min{k, ni}
λi =λ

so from dim N (λI − A)k for k = 1, 2, . . . we can determine the sizes of

the Jordan blocks associated with λ

Jordan canonical form 12–4

• factor out T and T −1, λI − A = T (λI − J)T −1

• for, say, a block of size 3:

   
0 −1 0 0 0 1
λiI−Ji =  0 0 −1  (λiI−Ji)2 =  0 0 0  (λiI−Ji)3 = 0
0 0 0 0 0 0

• for other blocks (say, size 3, for k ≥ 2)

 k k−1 k−2

(λi − λj ) −k(λi − λj ) (k(k − 1)/2)(λi − λj )
(λiI−Jj )k =  0 (λj − λi)k −k(λj − λi)k−1 
0 0 (λj − λi)k

Jordan canonical form 12–5

 Generalized eigenvectors

suppose T −1AT = J = diag(J1, . . . , Jq )

express T as
T = [T1 T2 · · · Tq ]
where Ti ∈ Cn×ni are the columns of T associated with ith Jordan block Ji
we have ATi = TiJi
let Ti = [vi1 vi2 · · · vini ]
then we have:
Avi1 = λivi1,
i.e., the first column of each Ti is an eigenvector associated with e.v. λi
for j = 2, . . . , ni,
Avij = vi j−1 + λivij

the vectors vi1, . . . vini are sometimes called generalized eigenvectors

Jordan canonical form 12–6

 Jordan form LDS

consider LDS ẋ = Ax

by change of coordinates x = T x̃, can put into form x̃˙ = J x̃

system is decomposed into independent ‘Jordan block systems’ x̃˙ i = Jix̃i

x̃ni x̃ni−1 x̃1

1/s 1/s 1/s

λ λ λ

Jordan blocks are sometimes called Jordan chains

(block diagram shows why)

Jordan canonical form 12–7

 Resolvent, exponential of Jordan block

resolvent of k × k Jordan block with eigenvalue λ:

 −1
s−λ −1
 s−λ ... 
(sI − Jλ) −1
=  ...

 −1 
s−λ
 
(s − λ) −1
(s − λ)
−2
· · · (s − λ)−k
 (s − λ)−1 · · · (s − λ)−k+1 
= 
 ... .. 

(s − λ)−1
= (s − λ)−1I + (s − λ)−2F1 + · · · + (s − λ)−k Fk−1

where Fi is the matrix with ones on the ith upper diagonal

Jordan canonical form 12–8

by inverse Laplace transform, exponential is:

tJλ tλ k−1


e = e I + tF1 + · · · + (t /(k − 1)!)Fk−1
 k−1

1 t · · · t /(k − 1)!

tλ  1 · · · tk−2/(k − 2)! 
= e  ... .. 

1

Jordan blocks yield:

• repeated poles in resolvent

• terms of form tpetλ in etA

Jordan canonical form 12–9

 Generalized modes

consider ẋ = Ax, with

x(0) = a1vi1 + · · · + ani vini = Tia

then x(t) = T eJtx̃(0) = TieJita

• trajectory stays in span of generalized eigenvectors

• coefficients have form p(t)eλt, where p is polynomial

• such solutions are called generalized modes of the system

Jordan canonical form 12–10

with general x(0) we can write

q
X
x(t) = etAx(0) = T etJ T −1x(0) = TietJi (SiT x(0))
i=1

where  
S1T
T −1 =  .. 
SqT

hence: all solutions of ẋ = Ax are linear combinations of (generalized)

modes

Jordan canonical form 12–11

 Cayley-Hamilton theorem

if p(s) = a0 + a1s + · · · + ak sk is a polynomial and A ∈ Rn×n, we define

p(A) = a0I + a1A + · · · + ak Ak

Cayley-Hamilton theorem: for any A ∈ Rn×n we have X (A) = 0, where

X (s) = det(sI − A)
 
1 2
example: with A = we have X (s) = s2 − 5s − 2, so
3 4

X (A) = A2 − 5A − 2I
   
7 10 1 2
= −5 − 2I
15 22 3 4
= 0

Jordan canonical form 12–12

corollary: for every p ∈ Z+, we have

p
 2 n−1

A ∈ span I, A, A , . . . , A

(and if A is invertible, also for p ∈ Z)

i.e., every power of A can be expressed as linear combination of

I, A, . . . , An−1

proof: divide X (s) into sp to get sp = q(s)X (s) + r(s)

r = α0 + α1s + · · · + αn−1sn−1 is remainder polynomial

then

Ap = q(A)X (A) + r(A) = r(A) = α0I + α1A + · · · + αn−1An−1

Jordan canonical form 12–13

for p = −1: rewrite C-H theorem

X (A) = An + an−1An−1 + · · · + a0I = 0

as

n−1
I = A −(a1/a0)I − (a2/a0)A − · · · − (1/a0)A
(A is invertible ⇔ a0 6= 0) so

A−1 = −(a1/a0)I − (a2/a0)A − · · · − (1/a0)An−1

i.e., inverse is linear combination of Ak , k = 0, . . . , n − 1

Jordan canonical form 12–14

 Proof of C-H theorem

first assume A is diagonalizable: T −1AT = Λ

X (s) = (s − λ1) · · · (s − λn)

since
X (A) = X (T ΛT −1) = T X (Λ)T −1
it suffices to show X (Λ) = 0

X (Λ) = (Λ − λ1I) · · · (Λ − λnI)

= diag(0, λ2 − λ1, . . . , λn − λ1) · · · diag(λ1 − λn, . . . , λn−1 − λn, 0)
= 0

Jordan canonical form 12–15

now let’s do general case: T −1AT = J

X (s) = (s − λ1)n1 · · · (s − λq )nq

suffices to show X (Ji) = 0

 ni
0 1 0 ···
X (Ji) = (Ji − λ1I)n1 · · ·  0 0 1 · · ·  · · · (Ji − λq I)nq = 0
...
| {z }
(Ji −λi I)ni

Jordan canonical form 12–16